INTERSPEECH 2004 - ICSLP
A Mel-LPC analysis is effective in speech recognition because of their auditory like frequency resolution. However, the spectral resolution is equal at all the frequency band. We proposed an Adaptive Mel-LPC Analysis Method (AMLPC). In AMLPC, the spectral resolution is difference according to phoneme category (vowel, fricative, etc). First, power normalized 1st-order auto-correlation r that represents the phoneme category is calculated from the input spectrum. Secondly, the frequency warped parameter "alpha" is calculated using r. The alpha is the coefficient of all-pass filter in MLPC, and is used to control the frequency resolution of spectral envelope. Finally, the warped predictors are obtained by each alpha parameter. The recognition performance of cepstrum parameters obtained by AMLPC was compared with that of the standard MLPC cepstrum through speaker independent word recognition. The results show that the AMLPC cepstrum leads to the improvement of error rate about 10% for the standard MLPC cepstrum.
Bibliographic reference. Nakatoh, Yoshihisa / Nishizaki, Makoto / Yoshizawa, Shinichi / Yamada, Maki (2004): "An adaptive MEL-LPC analysis for speech recognition", In INTERSPEECH-2004, 933-936.